422 big picture: Where are we
With reading papers?
Query
Planning
Deterministic Stochastic
• Value Iteration
• Approx. Inference
• Full Resolution
• SAT
Logics Belief Nets
Markov Decision Processes and
Partially Observable MDP
Markov Chains and HMMs First Order Logics
Ontologies
Applications of AI
Approx. : Gibbs
Undirected Graphical Models Markov Networks Conditional Random Fields
Reinforcement Learning Representation
Reasoning
Technique
Prob CFG Prob Relational Models Markov Logics
Hybrid: Det +Sto
Forward, Viterbi….
Approx. : Particle Filtering
CPSC 322, Lecture 34 Slide 1
11/9/2015 KCAP 2005 2
Extracting Knowledge from
Evaluative Text
Giuseppe Carenini, Raymond T. Ng, Ed Zwart
Computer Science Dept.
University of British Columbia
Vancouver, CANADA
11/9/2015 KCAP 2005 3
Motivation and Focus • Large amounts of info expressed in text form
is constantly produced
– News, Reports, Reviews, Blogs, Emails….
• Pressing need to extract and summarize
key/strategic info
Our Focus: evaluative text about single entity (good vs.
bad, right vs. wrong)
• Customer reviews
• Travel logs
• Job candidate evaluations…… etc.
• Considerable work but limited factual info
11/9/2015 KCAP 2005 4
KCAP from evaluative text (single entity)
• Extract relevant knowledge
• Summarize and present extracted knowledge
to user ………..
[Wilson et al. AAAI ’04]
[Hu, Liu KDD ’04]
[Hu, Liu AAAI ’04] [Popescu Etzioni HLT ’05]
A. What features of the entity are evaluated in the
reviews?
B. For each feature:
1. what is the polarity of the evaluation? (good vs. bad)
2. what is the strength of the evaluation? (rather good
vs. extremely good)
11/9/2015 KCAP 2005 5
Outline
• Feature Extraction: limitations of previous work
and our solution
• Evaluation of our approach
• Benefits in term of KCAP
• Conclusion and Demo of Future work
11/9/2015 KCAP 2005 6
…… the canon computer software used
to download , sort , . . . is very
nice and very easy to use. the only
two minor issues i have with the
camera are the lens cap ( it is not
very snug and can come off too
easily ). . . .
the menus are easy to navigate and
the buttons are easy to use. it is
a fantastic camera and well worth
the price .
Feature Extraction: sample form corpus [Hu&Liu 2004]
11/9/2015 KCAP 2005 7
Feature Extraction: sample form corpus [Hu&Liu 2004]
…… the canon computer software used
to download , sort , . . . is very
nice and very easy to use. the only
two minor issues i have with the
camera are the lens cap ( it is not
very snug and can come off too
easily ). . . .
the menus are easy to navigate and
the buttons are easy to use. it is
a fantastic camera and well worth
the price .
11/9/2015 KCAP 2005 8
Limitations of previous approach Key problems with extracted features (for KCAP):
• May be too many and redundant (often > 100)
• Flat, unstructured list (lack hierarchical organization)
• May be expressed in an unfamiliar terminology (for target user)
•spot metering
•metering option
•remote control
•battery
•night mode
•light automatic correction
•battery life
•remote
•battery charging system
•low light focus
•…
•battery life
•remote
•battery charging system
•low light focus
•…
•spot metering
•metering option
•remote control
•battery
•night mode
•light automatic correction
Lighting
Battery
11/9/2015 KCAP 2005 9
Example Ideal Mapping
1. Canon G3 PS Digital Camera [canon,canon PS g3, digital camera, camera,…]
1.1 User Interface [button, menus, lever]
1.2 Convenience [ ]
– Battery [battery life, battery charging system, battery]
– Self Timer [ ]
– Burst Mode [speed, mode]
– Rapid Fire Shot [speed]
– Delay Between Shots [unresponsiveness, delay, speed, lag time, lag]
– ….
2. Not Placed [manual, function, quality, strap, service, shoot, learning curve,…]
UDFs CFs
11/9/2015 KCAP 2005 10
Our Solution
• Map extracted features (Crude Features (CF)) in a
hierarchy of product features at different levels of
abstraction. Two alternatives:
– Learn the hierarchy
– Have the user provide a hierarchy of User Defined
Features (UDF)
• Such a mapping will:
– Eliminate redundancy (CFs with same referent
mapped in the same UDF)
– Provide conceptual structure
– Increase user familiarity with CFs
11/9/2015 KCAP 2005 11
Mixed-initiative Process
Corpus of
Evaluative Text
Merged Features
CF UDF UDFs
User can revise
mapping
User can revise
UDF
Mapping
method [Hu, Liu
AAAI ’04]
CFs
11/9/2015 KCAP 2005 12
Our Mapping Method • Map each CF in the “most similar” UDFs
• CFs and UDFs are terms (i.e., sequences of 1 or
more words)
So need a measure of term similarity
• Our approach to term similarity: combine
similarity between constituent words
• So need a measure of word similarity and a
function to combine similarities
11/9/2015 KCAP 2005 13
Word Similarity Metrics wm
• String Matching: baseline
• WordNet Synset Matching:
1 if the two words appear in the same synset….
e.g., (photograph, photo, exposure, picture )
• Wordnet Distance Matching: a set of measures that compute the semantic distance between the synsets of the two words
[Patwardan, Pedersen 2003] Cpan module
11/9/2015 KCAP 2005 14
Term similarity: Combine word scores
• MAX: terms' score is the maximum score of
comparing all possible word pairs
},...,{};,...,{ 11 mn wwudfvvcf
)},({max.
ji
ji
wvwm
• AVG: terms' score is the average of the max of all i with j,
and vice versa (to avoid a high score of one word dominate the
whole term's score)
2/
)},({max)},({max11
m
wvwm
n
wvwmm
j
ji
i
n
i
ji
j
11/9/2015 KCAP 2005 15
Mapping Algorithm Algorithm
• Each CF is mapped to the UDF with which it
receives the greatest similarity score
• In case of tie scores CF is mapped more than once
• But mapping occurs only if score greater than a
given threshold
Threshold
• For string and synset matching the threshold
was set to 0.
• For Wordnet distance similarity measures was
set by varying a parameter θ
11/9/2015 KCAP 2005 16
Outline
• Feature Extraction: limitations of previous work
and our solution
• Evaluation of our approach
• Benefits in term of KCAP
• Conclusion and Demo of Future work
11/9/2015 KCAP 2005 17
Results DigCam for AVG
Word
Similarity
metrics
Wordnet distance
(lin)
Mapping
quality
measures
↑ ↓
11/9/2015 KCAP 2005 18
Results DVD for AVG
Word
Similarity
metrics
Mapping
quality
measures
Wordnet distance
(lin)
↑ ↓
11/9/2015 KCAP 2005 19
Outline
• Feature Extraction: limitations of previous work
and our solution
• Evaluation of our approach
• Benefits in term of KCAP
• Conclusion and Demo of Future work
11/9/2015 KCAP 2005 20
Benefits in term of KCAP
CFs only
Key questions for manufactures and potential customers
- what product features are more frequently mentioned?
- how do customers evaluate those features?
- do they agree?
11/9/2015 KCAP 2005 21
Benefits in term of KCAP
CF UDF mapping
… answer the same questions
- different levels of abstraction
- less redundancy
- more familiar terms
image
picture
shot
11/9/2015 KCAP 2005 22
Conclusions
• Novel approach to feature extraction step in
KCAP from evaluative text
• Mixed-initiative mapping of flat list of extracted
CF into a UDF hierarchy
• Term similarity metrics
• Evaluation of these metrics on two corpora of
customer reviews : reasonable accuracy,
substantial reduction in redundancy
• Beneficial in term of captured knowledge
11/9/2015 KCAP 2005 23
Future Work
• Improve mapping method
– Try other term similarity measures (corpus based)
– Inject more sophisticated NLP (e.g., weight scoring
considering headword of a term)
• Summarize and present extracted knowledge to
user ……….. Combine text and graphics….
Adapt techniques for generating evaluative text
• Develop interface to support user revision of the
mapping and of the UDF hierarchy
Questions 2015-2 • Is WordNet the best online lexical database?!?
• Who is the user?
• UDFs / CFs / Gold Standard
• Unplaced CFs
• CF extraction and polarity (how many methods?)
• Constructing large UDF
• Different Languages
• Threshold
• Future
– Microsof Research took this over in 2007-8
– Interactive Multimedia Summarization (Visualization)
– Lexical Similarity vs. corpus-based
– Automatically create UDFs: Extract Hierarchy from the reviews/ from
existing ontologies - Speech input… Sarcasm 11/9/2015 KCAP 2005 24
11/9/2015 KCAP 2005 26
Data and Gold Standard
Two products: Digital Camera and DVD
• CFs from Hu&Liu annotated corpora: 101 CFs
for digital camera, 116 for DVD
• UDFs developed by domain experts: 86 UDFs
for digital camera, 38 for DVD
Gold Standard Development:
• We manually developed initial mappings
• User study: we asked 7 subjects to fix our
mappings with some random errors
• Based on their input a final version was created
11/9/2015 KCAP 2005 27
Measures of mapping quality
• (Graphical “distance” from correct placement)
))(()(_ ii cfedgeCountavgcfeistancdplacement
Zoom lever
Optical Zoom Lens
Manual Features Camera
● (Fraction of redundant CF's)
CF
FnonEmptyUDplacedCFreducredun
_
Can be maximized by placing all CFs in one UDF but….
redun_reduc in Gold Stand. DCam = .45 ; DVD=.43
11/9/2015 © Giuseppe Carenini 29
Talk Summary
Corpus of
Evaluative
Documents
Extract evaluative info:
feature hierarchy annotated
with evaluations
Generate
NL Summary
Multimedia Summary
Present
As Treemaps
Present
NL Summary Allow access to
original sources:
•Text footnotes
•Treemap zooming
Interactive [IUI ’06]
[EACL ’06]
[KCAP ’05]
SEA (NLG abstractor)
MEAD* (extractor)
[INLG ’08]
[IUI ’09]
• G. Carenini , J. Cheung , A. Pauls. Multi-Document Summarization of
Evaluative text, Computational Intelligence, 2012
• Carenini G. and Rizoli L., A Multimedia Interface for Facilitating
Comparisons of Opinions. . In: Proceedings of the 13th International
Conference on Intelligent User Interfaces, (IUI 2009), Sanibel Island,
Sydney, Florida, USA, 2009 [pdf]
• Carenini G, Cheung J., Extractive vs. NLG-based Abstractive
Summarization of Evaluative Text: The Effect of Corpus Controversiality.
International Conference on Natural Language Generation. (INLG 2008),
Salt Fork, Ohio, USA, June 12-14, 2008 [pdf]
11/9/2015 KCAP 2005 30
Carenini G., Ng R., Pauls A. MultiDocument Summarization of
Evaluative Text, Proceedings of the 11th European Chapter of the
Association for Computational Linguistics (EACL 2006), Trento, Italy,
April, 2006. [pdf]
Carenini G., Ng R., Pauls A. Interactive Multimedia Summaries of
Evaluative Text, Proceedings of the 10th International Conference on
Intelligent User Interfaces (IUI 2006), Sydney, Australia, Gen29-Feb1,
2006. [pdf]
11/9/2015 © Giuseppe Carenini 31
Multimedia Interactive Approach
Corpus of
Evaluative
Documents
Extract
evaluative info
Generate NL
Summary
Multimedia Summary
Present
in Graphics
Present
NL Summary Allow access to
Original sources
Interactive
11/9/2015 © Giuseppe Carenini 32
Extracted evaluative info after mapping
Canon G3 PS Digital Camera [-1,-1,+1,+2,+2,+3,+3,+3]
1. User Interface [ +2]
– Button [ +1]
– Menus [+2,+2,+2,+3+3]
– Lever [ ]
2. Convenience [ ]
– Battery [ ]
• Battery life [-1,-1,-2 ]
• Battery charging system [ ]
– ….
3. ….
• Merged Features hierarchy annotated with all the
evaluations each feature received in the corpus
PSi is the set of polarity/strength
evaluations for feature fi
11/9/2015 © Giuseppe Carenini 33
Conveying extracted info with graphics
Treemaps: space-filling technique for visualizing hierarchical information structures
• Each node in the hierarchy is represented as a rectangle
• Descendants of a node are represented as nested rectangles
• Rectangle size and colour can express information about the node
Visualization should convey:
• Hierarchical organization of the features
• For each feature
– # of evaluations
– polarity/strength of the evaluations (average?)
11/9/2015 © Giuseppe Carenini 35
One possibleTreemap
• Each product feature is represented as a rectangle
• The hierarchy is represented by nesting
• Rectangle size expresses # of evaluations
• Rectangle colour expresses avg polarity/strength of evaluations (black for neutral, the more positive/negative the more green/red)
11/9/2015 © Giuseppe Carenini 36
Apex DVD Player Extra Features Disk Format
Video Output User Interface
11/9/2015 © Giuseppe Carenini 37
Another possible Treemap
• Each evaluation is represented as a rectangle
• The hierarchy is represented by nesting
• Rectangle colour expresses polarity/strength of the evaluation (black for neutral, the more positive/negative the more green/red)
• Note: More effective in conveying controversiality
11/9/2015 © Giuseppe Carenini 40
Multimedia Interactive Approach
Corpus of
Evaluative
Documents
Extract
evaluative info
Generate NL
Summary
Multimedia Summary
Present
in Graphics
Present
NL Summary Allow access to
Original sources
Interactive
11/9/2015 © Giuseppe Carenini 43
Multimedia Interactive Summary:
Formative Evaluation
• Procedure (similar to study-1 and study-2)
• Interested in testing effectiveness of text graphics combination (redundancy / support)
• Very positive feedback (Details in IUI-06 paper)
• Recent Extension to comparison of two entities (see paper in IUI-09)
11/9/2015 © Giuseppe Carenini 44
Talk Summary
Corpus of
Evaluative
Documents
Extract evaluative info:
feature hierarchy annotated
with evaluations
Generate
NL Summary
Multimedia Summary
Present
As Treemaps
Present
NL Summary Allow access to
original sources:
•Text footnotes
•Treemap zooming
Interactive [IUI ’06]
[EACL ’06]
[KCAP ’05]
SEA (NLG abstractor)
MEAD* (extractor)
[INLG ’08]
[IUI ’09]
Questions 2015 • UDFs / CFs / Gold Standard
• Unplaced CFs
• Clarification Placement distance
• CF extraction and polarity
• Constructing large UDF
• Different Languages
• Trade-off Placement and Redundancy
• Future
– MSR
– Interactive Multimedia Summarization
– Extract Hierarchy from the reviews (automatically create
UDFs)….. Speech input… Sarcasm 11/9/2015 KCAP 2005 46
Placement Distance
• The accuracy of a CF term in the research is
assessed by considering the hierarchical path
distance between where it is placed by the
mapping algorithm and where it is placed by
the gold standard (control
mapping). Does the research assume that
path lengths all encode the same semantic
distance? (e.g. that pixels (parent) - >
resolution (child) has a semantic subset
distance equal to image (parent) -> image type
(child)) 11/9/2015 KCAP 2005 47
• G. Carenini , J. Cheung , A. Pauls. Multi-Document Summarization of
Evaluative text, Computational Intelligence, 2012
• Carenini G. and Rizoli L., A Multimedia Interface for Facilitating
Comparisons of Opinions. . In: Proceedings of the 13th International
Conference on Intelligent User Interfaces, (IUI 2009), Sanibel Island,
Sydney, Florida, USA, 2009 [pdf]
• Carenini G, Cheung J., Extractive vs. NLG-based Abstractive
Summarization of Evaluative Text: The Effect of Corpus Controversiality.
International Conference on Natural Language Generation. (INLG 2008),
Salt Fork, Ohio, USA, June 12-14, 2008 [pdf]
11/9/2015 KCAP 2005 48
Carenini G., Ng R., Pauls A. MultiDocument Summarization of
Evaluative Text, Proceedings of the 11th European Chapter of the
Association for Computational Linguistics (EACL 2006), Trento, Italy,
April, 2006. [pdf]
Carenini G., Ng R., Pauls A. Interactive Multimedia Summaries of
Evaluative Text, Proceedings of the 10th International Conference on
Intelligent User Interfaces (IUI 2006), Sydney, Australia, Gen29-Feb1,
2006. [pdf]
11/9/2015 KCAP 2005 51
Support Analysis of Evaluative
Arguments (about single entity)
Corpus of
relevant
Evaluative
Arguments
NLG
Summary
Multimedia Interactive
summary
Merge Features
+ strength + polarity
(for each evaluation)
NLG-SEA
KEEA
Treemap
Engine
Value treemap
SE-SEA
SentExt
Summary E
E
UDF
AMVF
MISEA
11/9/2015 KCAP 2005 53
From Mapping To Evaluation
• Given unsupervised CF extraction and
unsupervised UDF<->CF mapping, need
to evaluate UDF features
• Assume we can calculate strength and
polarity of customer evaluations for each
CF using existing methods (Hu & Liu
2004; Wilson et al. 2004),
– then we can generate an evaluation for each
UDF based on its CF's
11/9/2015 KCAP 2005 55
Plan for Summary Generation
• Adapt GEA (Generator of Evaluative
Arguments) (Carenini & Moore 2001) for
– Content selection and organization
– Microplanning (partially)
– Realization
• Adapt existing MEAD (Radev et. al. 2001)
software as baseline “domain/task
independent” summarizer
• Evaluation: Compare system against
baseline with human judges
11/9/2015 KCAP 2005 56
Generator of Evaluative
Arguments (GEA)
• Generates evaluations of entities based on:
– properties of entity
– user preferences about that entity
• Entity is represented as a set of attributes
and values (e.g. (Zoom range . 12x))
• User Preferences are modelled using an
AMVF (Additive Multiattribue Value
Function)
– This is a hierarchical set of preferences about
entity, with attributes as leafs
11/9/2015 KCAP 2005 57
GEA example: AMVF
Location
Amenities
Porch-Size
Deck-Size
0.7
0.4
0.6
0.3
0.2
0.8
Neighborhood
Park-Distance House Value
11/9/2015 KCAP 2005 58
GEA example: Attributes
Location
Amenities
Porch-Size
Deck-Size
0.9
0.25
0.6
Neighborhood
Park-Distance
House-A
n2
0.5 km
20 m2
36 m2
House Value
0.6 +
+
+
_
_ + Likes it
Does not like it
Domain Values
Component Value
Function
Domain Values Domain Value Attribute
Evaluation
11/9/2015 KCAP 2005 59
GEA example: Opposing/Supporting Evidence
Location
Amenities
Porch-Size
Deck-Size
0.7
0.4
0.6
0.3
0.2
0.8
0.64 0.9
0.25
0.6
0.32
0.78
Neighborhood
Park-Distance
House-A
n2
0.5 km
20 m2
36 m2
House Value
0.6
_
+
_
+
+
+
_
+ Supporting
Opposing
+
_
+
+
+
+
_
_ + Likes it
Does not like it
opposing
opposing
supporting
supporting
relation
+ _
+
+
+
_
_
_
o Parent(o)
11/9/2015 KCAP 2005 60
Measure of Importance [Klein 94]
_
+ Supporting
Opposing
Location
Amenities
Porch-Size
Deck-Size
0.7
0.4
0.6
0.3
0.2
0.8
0.64 0.9
0.25
0.6
0.32
0.78
Neighborhood
Park-Distance
House-A
n2
0.5 km
20 m2
36 m2
House Value
0.6
+
_
_
+
+
+
+
_
+
+
+
+
_
_ + Likes it
Does not like it
For each attribute a :
Importance a wamax va , 1 va
0.55
0.2
0.12
0.6
0.24
0.54
11/9/2015 KCAP 2005 61
Argumentative Strategy
Selection: include only “important” evidence
(i.e., above threshold on measure of importance)
Organization:
(1) Main Claim (e.g., “This house is interesting”)
(2) Opposing evidence
(3) Most important supporting evidence
(4) Further supporting evidence -- ordered by importance with strongest last
Strategy applied recursively on supporting evidence
Based on guidelines from argumentation theory [Miller 96, Mayberry 96]
11/9/2015 KCAP 2005 62
Adapting GEA
GEA
• AMVF hierarchy ->
• AMVF weights ->
• Component ->
Value Function
Customer Reviews ● UDF hierarchy ● Relative frequency
of UDFs in corpus ● Aggregation of
polarity/strength of UDF features
11/9/2015 KCAP 2005 63
Adapting GEA (cont'd)
• Differences
– Customers may evaluate non-leaf elements
(e.g. “Location”) directly
– in GEA domain, entities had only one
evaluation for each attribute
• For customer reviews, must give some
indication of distribution of customer opinions
on each attribute
11/9/2015 KCAP 2005 64
Example: Some (fake) Reviews
“I really liked the Canon G3[+2]. The 12x zoom is
really useful[+1]! The only thing I didn't like
was its poor [-1] focussing in low light.”
“The Canon G3 is a great deal. The lens features
were the best I've seen for a camera of its
price[+2]. The menu system is very
intuitive[+1], but I wish the camera could take
RAW images[-1].”
“I really didn't like this camera[-2]. It
focussed very poorly [-2] indoors (when I use
it most) and I found myself wishing there were
more modes on the dial [-1] rather than in the
menu system. I returned mine already.”
11/9/2015 KCAP 2005 65
Adapted GEA
Lens
Interface
Dial
Menu
0.57
0.25
0.5
0.43
0.66
0.33
0.47 0.25
0.83
0.25
0.39
0.5
Zoom
Auto-Focus
Canon G3
+1
-1, -2
+2
-1,-2
Canon G3
0.66
+
_
_
+
+
+
+
_
+
+
+
_
0.29
0.26
0.495
0.28
0.17
0.69
+2
_
Strength/Polarity
of User
Evaluations
Attribute
Evaluation
Evaluation Aggregation
Function
11/9/2015 KCAP 2005 66
Output of GEA
• What GEA gives us:
– High-level text plan (i.e. content selection and
ordering)
– Cue phrases for argumentation strategy (“In fact”,
“Although”, etc.)
• What GEA does not give us:
– Appropriate micro-planning (lexicalization).
• Need to give indication of distribution of customer
opinions
11/9/2015 KCAP 2005 67
Hypothetical GEA Output
T h e C a n o n G 3 i s a g o o d c a m e r a . H o w e v e r , t h e
i n t e r f a c e f e a t u r e i s p o o r. A l t h o u g h t h e m e n u
s y s t e m i s g o o d , t h e d i a l s y s t e m i s t e r r i b l e .
11/9/2015 KCAP 2005 68
Target Summary
M o s t u s e r s t h o u g h t C a n o n G 3 w a s a g o o d c a m e r a .
H o w e v e r , s e v e r a l u s e r s d i d n o t l i k e
i n t e r f a c e . A l t h o u g h m o s t u s e r s l i k e d t h e m e n u
s y s t e m , m a n y t h o u g h t t h e d i a l w a s t e r r i b l e .
11/9/2015 KCAP 2005 69
Evaluation
• Current idea: task-based (extrinsic)
evaluation
– Give human test subject summary
– Then, allow user some fixed time (e.g. 5
minutes) to scan a corpus of reviews (20-30?)
– User should then answer e.g.
• if summary provides “all” (?) important information
• if summary left out information
• if missing information was important
• if summary is representative of corpus
– Also evaluate fluency with known methods
– Others?
11/9/2015 KCAP 2005 70
• Current method of adapting GEA is just a first
pass.
– Could change e.g. Measure of Importance.
• We may leverage GEA's ability to create user-
tailored evaluative arguments for generating
user-tailored summaries (long term)
Future Directions
11/9/2015 KCAP 2005 71
IE Key Sub-tasks
A. What features of the objects are evaluated
in the reviews?
B. For each feature:
i. what is the polarity of the evaluation? (good
vs. bad)
ii. what is the strength of the evaluation? (rather
good vs. extremely good)
11/9/2015 KCAP 2005 72
(User-Specific) Summarization of
Multiple Customer Reviews
The Goal: An automatic solution to the problem of summarizing a potentially large set of documents that contain evaluative language about a given entity (e.g., a product, a location, a job candidate, etc.) User Specific: the summary is tailored to user’s conceptualization of the entity (now) model of the user’s preferences (long term)
11/9/2015 KCAP 2005 73
Example: Some (fake) Reviews
“I really liked the Canon G3. The 12x zoom is
really useful! The only thing I didn't like
was its poor focussing in low light.”
“The Canon G3 is a great deal. The lens
features were the best I've seen for a camera
of its price. The menu system is very
intuitive, but I wish the camera could take
RAW images.”
“I really didn't like this camera. It focussed
very poorly indoors (when I use it most) and
I found myself wishing there were more modes
on the dial rather than in the menu system. I
returned mine already.”
11/9/2015 KCAP 2005 74
Example: Target Summary
M o s t u s e r s l i k e d t h e C a n o n G 3 . M a n y f o u n d t h e
z o o m f e a t u r e t o b e g o o d. A l t h o u g h m a n y u s e r s
d i d n o t l i k e t h e a u t o f o c u s , a f e w u s e r s
l i k e d t h e m e n u s y s t e m . O n l y 1 u s e r d i d n o t
l i k e t h e c a m e r a .
11/9/2015 KCAP 2005 75
Example Target Summary
• Features
– Selection of content (flash range not mentioned)
– Discourse cues (cue phrases, order of evidence)
– Contrasting and supporting evidence for summary of
camera
– Lexicalization of numerical tallies (2/3 =>
“most”)
11/9/2015 KCAP 2005 77
Example of Learned Features for a
Digital Camera
• noise
• function
• button
• camera
• four megapixel
• remote control
• software
• manual
• remote
• lever
• price
• Canon G3
• strap
• low light focus
• tiff format
• use
11/9/2015 KCAP 2005 78
Ideal Extraction: sample form corpus [Hu&Liu 2004]
…… the canon computer software [+2]
used to download , sort , . . . is
very nice and very easy to use. the
only two minor issues i have with
the camera are the lens cap [-1] (
it is not very snug and can come
off too easily ). . . .
the menus [+1] are easy to navigate
and the buttons [+1] are easy to
use. it is a fantastic camera [+3]
and well worth the price .